[PATCH 07/11] fsmc/nand: Provide contiguous buffers to dma

[Linus Walleij]

On Tue, Oct 9, 2012 at 12:44 PM, Vipin Kumar <vipin.kumar at st.com> wrote:

> read_buf/write_buf callbacks should be able to accept a user space memory
> address (virtually contiguous memory) as buffer pointer.
>
> This patch allocates a logically contiguous memory area which is use for dma

You mean PHYSICALLY contigous, don't you?

> xfers during read and write accesses.
>
> Signed-off-by: Vipin Kumar <vipin.kumar at st.com>

If you really want a physically contigous buffer you need to use
CMA, but I don't think that is the real problem here...

We're already using userspace buffers in e.g. the MMCI driver
(drivers/mmc/host/mmci.c).

The real problem is likely the DMA driver. The stuf that get
fed into dma.device_prep_dma_memcpy() needs to be
converted to a scatterlist and then set up in the LLI list
for the controller.

IIRC SPEAr is using drivers/dma/dw_dmac.c so
check this driver's dwc_prep_dma_memcpy().
It does seem like it is checking whether src or
dest is scattered in this for() loop:

for (offset = 0; offset < len; offset += xfer_count << src_width) {}

dma_sync_single_for_device() is translating the virtual
address to physical for every chunk BTW.

So instead of doing this copying, debug the problem, and
see if there is a bug in that for()-loop or similar, if it needs
to be rewritten or so.

Yours,
Linus Walleij